Resolves a conflict in model/pipeline.py: git's line-based merge would
have silently reverted the transformer quantization from
Float8DynamicActivationFloat8WeightConfig() back to main's stale
Float8WeightOnlyConfig() (a #35 stopgap that predates the #39 AOT-compile
fix making dynamic-activation quantization safe again). The precompiled
AOTI package loaded via spaces.aoti_load() expects the dynamic-activation
op graph, so that silent revert would have mismatched at runtime.
Resolved by keeping develop's dynamic-activation config.